Internal Simulation: History and Intuition of World Models

A friend throws a tennis ball from three meters away. Before you have even registered the ball's spin, your hand has already moved to the right position. This is neither reflex nor luck: your brain simulated the ball's flight path in that fraction of a second, predicted where it would land, and directed your muscles to act in advance.

A world model is the computational version of this mechanism: a sketch of "how the world works" maintained internally by a brain or an AI, allowing an agent to mentally rehearse outcomes before acting in reality.

This lecture is organized in three parts:

• Conceptual foundations: from Craik (1943) to predictive coding, the internal model principle, and the boundary between broad and narrow definitions of world models
• Four eras: the evolutionary arc from RNN beginnings to Ha & Schmidhuber, Dreamer, and JEPA
• Why now: the converging context of video generation, embodied intelligence, and autonomous driving, plus the course roadmap